Why autonomous driving requires so many sensors

When we drive a car, our brain has to discern everything that happens ahead of and around us, depending on what our senses perceive and on the knowledge we have of our route, or on the intentions we have when travelling in the vehicle.

All of that and much more is what autonomous driving has to do in order to be even safer than a human being when moving a car without anyone intervening. For this reason, manufacturers of autonomous cars always boast about the high number of sensors, cameras, ultrasound systems, radars or LIDAR-type equipment they incorporate into their models. The more of these elements a car carries on board, the better. And why better? Just look at the picture accompanying these lines to understand it.

The photograph shows what technicians working on the development of autonomous vehicles know as a “corner case”. These cases are those in which a vehicle will behave unpredictably because its software has processed an unusual scenario differently from how a human being would. The image shows the data recorded by the car’s sensors and camera that led the autonomous vehicle to the mistaken conclusion that the images of the cyclists at the rear of the minivan ahead are real human cyclists, when in fact it is only a photograph.

At this blind spot, “the LIDAR cannot detect glass, radar mainly distinguishes metal, and the camera can be fooled by images,” explains Danny Atsmon, CEO of Cognata, a company that develops software simulators that expose vehicles to corner cases so that engineers can develop the algorithms needed for autonomous vehicles to handle them without risking accidents. According to Atsmon, “Each of the sensors needed for autonomous driving solves a different part of the detection challenge.” By gradually discovering which data can be used to correctly resolve specific corner cases (both through simulations and in real life), cars learn to deal with increasingly complex situations.

Tesla was criticised for its decision to rely only on radar, cameras and ultrasonic sensors to provide data to its Autopilot system after one of its vehicles failed to distinguish a trailer from the bright blue sky. This error caused a collision between the vehicles and killed the driver of the Tesla vehicle instantly. Critics argue that LIDAR is a key element in the range of sensors. Unlike a camera, this device works well in low light and glare, and provides more detailed data than radar or ultrasound. But, as Atsmon points out, “even LIDAR has flaws“: for example, it cannot distinguish between a red and a green traffic light.

Hence, the safest bet for avoiding these situations is for manufacturers to keep relying on a greater diversity of sensors with a view to creating the greatest possible redundancy in their systems. This is why the largest possible number of detectors, sensors and cameras must be installed on board all autonomous cars. Thanks to them, most autonomous vehicles overcome problems like the one posed by both the photograph and the video images we offer, where an autonomous car is unable to react to overcome the prohibition that, for it, crossing the solid line of the circle drawn around it represents.

Combining and analysing information from all the on-board detection systems with stored experience is what will allow it to overcome these difficulties. The cyclists, at least, will surely be grateful.

https://vimeo.com/208642358